Partial atomic charges provide an intuitive and efficient way to describe the charge distribution and the resulting intermolecular electrostatic interactions in liquid water. Many charge models exist and it is unclear which model provides the best assignment of partial atomic charges in response to the local molecular environment. In this work, we systematically scrutinize various electronic structure methods and charge models (Mulliken, Natural Population Analysis, CHelpG, RESP, Hirshfeld, Iterative Hirshfeld, and Bader) by evaluating their performance in predicting the dipole moments of isolated water, water clusters, and liquid water as well as charge transfer in the water dimer and liquid water. Although none of the seven charge models is capable of fully capturing the dipole moment increase from isolated water (1.85 D) to liquid water (about 2.9 D), the Iterative Hirshfeld method performs best for liquid water, reproducing its experimental average molecular dipole moment, yielding a reasonable amount of intermolecular charge transfer, and showing modest sensitivity to the local water environment. The performance of the charge model is dependent on the choice of the density functional and the quantum treatment of the environment. The computed molecular dipole moment of water generally increases with the percentage of the exact Hartree-Fock exchange in the functional, whereas the amount of charge transfer between molecules decreases. For liquid water, including two full solvation shells of surrounding water molecules (within about 5.5 A of the central water) in the quantum-chemical calculation converges the charges of the central water molecule. Our final pragmatic quantum-chemical charge assigning protocol for liquid water is the Iterative Hirshfeld method with M06-HF/aug-cc-pVDZ and a quantum region cutoff radius of 5.5 A.<br>
Machine Learning Inference of Molecular Dipole Moment in Liquid Water
